RESEARCH ARTICLE
Predicting human protein subcellular localization by heterogeneous and comprehensive approaches Chi-Hua Tung1, Chi-Wei Chen2, Han-Hao Sun2, Yen-Wei Chu2,3* 1 Department of Bioinformatics, Chung-Hua University, Hsinchu, Taiwan, 2 Institute of Genomics and Bioinformatics, National Chung Hsing University 250, Taichung 402, Taiwan, 3 Biotechnology Center, Agricultural Biotechnology Center, Institute of Molecular Biology, Graduate Institute of Biotechnology, National Chung Hsing University 250, Taichung 402, Taiwan * [email protected]
a1111111111 a1111111111 a1111111111 a1111111111 a1111111111
OPEN ACCESS Citation: Tung C-H, Chen C-W, Sun H-H, Chu Y-W (2017) Predicting human protein subcellular localization by heterogeneous and comprehensive approaches. PLoS ONE 12(6): e0178832. https:// doi.org/10.1371/journal.pone.0178832 Editor: Bin Liu, Harbin Institute of Technology Shenzhen Graduate School, CHINA Received: March 8, 2017
Abstract Drug development and investigation of protein function both require an understanding of protein subcellular localization. We developed a system, REALoc, that can predict the subcellular localization of singleplex and multiplex proteins in humans. This system, based on comprehensive strategy, consists of two heterogeneous systematic frameworks that integrate one-to-one and many-to-many machine learning methods and use sequence-based features, including amino acid composition, surface accessibility, weighted sign aa index, and sequence similarity profile, as well as gene ontology function-based features. REALoc can be used to predict localization to six subcellular compartments (cell membrane, cytoplasm, endoplasmic reticulum/Golgi, mitochondrion, nucleus, and extracellular). REALoc yielded a 75.3% absolute true success rate during five-fold cross-validation and a 57.1% absolute true success rate in an independent database test, which was >10% higher than six other prediction systems. Lastly, we analyzed the effects of Vote and GANN models on singleplex and multiplex localization prediction efficacy. REALoc is freely available at http:// predictor.nchu.edu.tw/REALoc.
Accepted: May 19, 2017 Published: June 28, 2017 Copyright: © 2017 Tung et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: All relevant data are within the paper and its Supporting Information files. Funding: This research was supported by Ministry of Science and Technology, Taiwan, R.O.C. under grant number 105-2221-E-216-021. Competing interests: The authors have declared that no competing interests exist.
Introduction Current efforts in genomics science routinely make use of various fast and accurate sequencing platforms. These technologies have led to greater understanding of protein function and the regulation of biological networks, and the rapidly accumulated sequence information has allowed exploration of complex physiological mechanisms and numerous diseases. Protein localization in cells is often closely correlated with its function. According targeting signals may occur anywhere in the protein sequence, proteins will be sorted to their destination. For example, signal peptide is a short N-terminal amino acid sequence that guide the distribution of the protein to the membrane of the endoplasmic reticulum (ER) and enter the secretory pathway [1,2]. Therefore, we endeavored to use the available sequence information to provide quick subcellular localization prediction. Two transduction pathways exist: Co-translational
PLOS ONE | https://doi.org/10.1371/journal.pone.0178832 June 28, 2017
1 / 14
Realoc
translocation involves signal-recognition particles and the endoplasmic reticulum delivery system, whereas in post-translational translocation, translation is completed in the cytoplasm, but the protein retains information specifying delivery to different organelles [3]. Nakai et al. (1991) first proposed the prediction of subcellular localization via signal peptides. However, due to the very recent growth of protein databases and development of different machine learning methods, TargetP and SignalP 4.0 are currently the only commonly known prediction systems [4,5], yet they can only predict localization to three subcellular compartments (chloroplast, mitochondrion, and extracellular). Methods for predicting subcellular localization can be classified into the following three types: 1) Homology-based prediction compares the localization of known proteins with unknown proteins. If a certain degree of similarity is found in the sequence, then it can be inferred that the unknown protein’s subcellular localization may be the same as the known protein [6]. The Basic Local Alignment Search Tool (BLAST) sequence alignment method [7] has been widely applied in several prediction systems, including WegoLoc, iLoc-Hum, and Euk-mPloc 2.0 [8–10]. However, when the similarity between the unknown protein and the database is low, this method has poor predictive ability. Hence, it is often integrated with other methods. 2) Functional domain-based prediction relies on known structures or functional data, such as protein functional domains and motifs, as well as information in the gene ontology (GO) database [6,8,11–14]. There are many learning models of research methods are used to establish the relevance of GO terms and subcellular localization. [10,12,15–19] It has been shown that GO terms can be used to advance the performance of subcellular localization prediction. These functional data regarded as domain knowledge are highly accurate and reliable, but this approach requires manual verification of each annotation and cannot be applied to the whole new protein; therefore it is usually combined with the homology-based approach. The optimized combined approach can greatly increase the predictive accuracy. 3) Sequence-based prediction relies on information about the primary amino acid sequence of proteins used for information technology operations or discovery of hidden information, commonly including amino acid composition, pseudo amino acid composition, and n-grams [19–22]. Prediction results from this approach are typically less informative than those for homology- and functional domain-based methods, but in predicting subcellular localization of unknown proteins, it is still a feasible approach. There are also systems based on n-grams for processing the amino acid sequence to analyze the subcellular location of the target protein [21]. Most of the existing methods described above only predict singleplex proteins, those that localize to a single subcellular location, and they do not consider multiplex proteins, those that localize to two or more different compartments within the cell. Identifying multiple locations of a multiplex protein has high value to understanding biological functions, and there is much room for continued development in this area [23]. Currently, the development of subcellular location prediction tools faces two major problems: the difference among the data quantity of locations is too large and the poor predictive ability for multiplex proteins [24]. Therefore, a highly accurate system was developed for predicting human protein subcellular localization that we called REALoc (Reliable and Effective methods to Assist predicting human protein subcellular Localization), which consists of two systematic frameworks. The first level uses 32 support vector machine (SVM) models in a oneto-one relationship system to completely cover all negative learning information by comprehensive strategy; the second level uses a many-to-many relationship system comprised of two learning mechanisms, a genetic algorithm optimized neural network (GANN) and majority voting (Vote), which represent the strong and weak data correlation, respectively. REALoc can predict localization to the cell membrane, cytoplasm, endoplasmic reticulum/Golgi, mitochondrion, nucleus, and extracellular. With regards to the learning sequences, this study adopted
PLOS ONE | https://doi.org/10.1371/journal.pone.0178832 June 28, 2017
2 / 14
Realoc
not only the commonly used amino acid composition (AAC) and surface accessibility (SA), but also two new learning features, weighted sign amino acid index (aa index) and sequence similarity profile. Weighted sign aa index allows more emphasis to be placed on subcellular localization-related amino acid features in the learning process, while sequence similarity profile provides identifiable learning information for similar sequences in proteins with different subcellular localization. In regards to structural learning features, we used the regular maximum relevance minimum redundancy (regular-mRMR) method to select the 35 best localization- related sets of features for GO. An independent testing dataset was used to verify the predictive power of REALoc, and the prediction results were compared to six current predictors, CELLO [25], locTree2 [26] iLochum [14], Hum-mPLoc 2.0 [27], GOASVM [6] and mGOF-loc [22]. REALoc also uses the absolute true success rate (ATSR) [9,14,24] to represent the protein prediction score, which avoids under- and over-prediction. Using the systematic ATSR evaluation method we objectively evaluated the reliability of the human subcellular localization prediction system. After five-fold cross-validation, the ATSR with the training data and independent testing data used in this study were 75.3 and 56.5%, respectively. Whether a protein was located in a single or multiple subcellular compartments, REALoc showed higher overall performance compared with other prediction systems, with a >10% improvement in the ATSR in independent testing. In terms of the analysis of the REALoc system, we also removed the GO information to evaluate the predictive power for novel proteins, and the prediction efficacies of the Vote and GANN learning mechanisms for singleplex and multiplex protein data were also analyzed.
Materials and methods Data processing Training dataset 5939p. The dataset was obtained from the UniProtKB/Swiss-Prot protein database October 2011 version [28]. Proteins with a comment on subcellular location were extracted and those with uncertain terms such as “by similarity,” potential,” and “probable” were removed. Sequences
Predicting human protein subcellular localization by heterogeneous and comprehensive approaches Chi-Hua Tung1, Chi-Wei Chen2, Han-Hao Sun2, Yen-Wei Chu2,3* 1 Department of Bioinformatics, Chung-Hua University, Hsinchu, Taiwan, 2 Institute of Genomics and Bioinformatics, National Chung Hsing University 250, Taichung 402, Taiwan, 3 Biotechnology Center, Agricultural Biotechnology Center, Institute of Molecular Biology, Graduate Institute of Biotechnology, National Chung Hsing University 250, Taichung 402, Taiwan * [email protected]
a1111111111 a1111111111 a1111111111 a1111111111 a1111111111
OPEN ACCESS Citation: Tung C-H, Chen C-W, Sun H-H, Chu Y-W (2017) Predicting human protein subcellular localization by heterogeneous and comprehensive approaches. PLoS ONE 12(6): e0178832. https:// doi.org/10.1371/journal.pone.0178832 Editor: Bin Liu, Harbin Institute of Technology Shenzhen Graduate School, CHINA Received: March 8, 2017
Abstract Drug development and investigation of protein function both require an understanding of protein subcellular localization. We developed a system, REALoc, that can predict the subcellular localization of singleplex and multiplex proteins in humans. This system, based on comprehensive strategy, consists of two heterogeneous systematic frameworks that integrate one-to-one and many-to-many machine learning methods and use sequence-based features, including amino acid composition, surface accessibility, weighted sign aa index, and sequence similarity profile, as well as gene ontology function-based features. REALoc can be used to predict localization to six subcellular compartments (cell membrane, cytoplasm, endoplasmic reticulum/Golgi, mitochondrion, nucleus, and extracellular). REALoc yielded a 75.3% absolute true success rate during five-fold cross-validation and a 57.1% absolute true success rate in an independent database test, which was >10% higher than six other prediction systems. Lastly, we analyzed the effects of Vote and GANN models on singleplex and multiplex localization prediction efficacy. REALoc is freely available at http:// predictor.nchu.edu.tw/REALoc.
Accepted: May 19, 2017 Published: June 28, 2017 Copyright: © 2017 Tung et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: All relevant data are within the paper and its Supporting Information files. Funding: This research was supported by Ministry of Science and Technology, Taiwan, R.O.C. under grant number 105-2221-E-216-021. Competing interests: The authors have declared that no competing interests exist.
Introduction Current efforts in genomics science routinely make use of various fast and accurate sequencing platforms. These technologies have led to greater understanding of protein function and the regulation of biological networks, and the rapidly accumulated sequence information has allowed exploration of complex physiological mechanisms and numerous diseases. Protein localization in cells is often closely correlated with its function. According targeting signals may occur anywhere in the protein sequence, proteins will be sorted to their destination. For example, signal peptide is a short N-terminal amino acid sequence that guide the distribution of the protein to the membrane of the endoplasmic reticulum (ER) and enter the secretory pathway [1,2]. Therefore, we endeavored to use the available sequence information to provide quick subcellular localization prediction. Two transduction pathways exist: Co-translational
PLOS ONE | https://doi.org/10.1371/journal.pone.0178832 June 28, 2017
1 / 14
Realoc
translocation involves signal-recognition particles and the endoplasmic reticulum delivery system, whereas in post-translational translocation, translation is completed in the cytoplasm, but the protein retains information specifying delivery to different organelles [3]. Nakai et al. (1991) first proposed the prediction of subcellular localization via signal peptides. However, due to the very recent growth of protein databases and development of different machine learning methods, TargetP and SignalP 4.0 are currently the only commonly known prediction systems [4,5], yet they can only predict localization to three subcellular compartments (chloroplast, mitochondrion, and extracellular). Methods for predicting subcellular localization can be classified into the following three types: 1) Homology-based prediction compares the localization of known proteins with unknown proteins. If a certain degree of similarity is found in the sequence, then it can be inferred that the unknown protein’s subcellular localization may be the same as the known protein [6]. The Basic Local Alignment Search Tool (BLAST) sequence alignment method [7] has been widely applied in several prediction systems, including WegoLoc, iLoc-Hum, and Euk-mPloc 2.0 [8–10]. However, when the similarity between the unknown protein and the database is low, this method has poor predictive ability. Hence, it is often integrated with other methods. 2) Functional domain-based prediction relies on known structures or functional data, such as protein functional domains and motifs, as well as information in the gene ontology (GO) database [6,8,11–14]. There are many learning models of research methods are used to establish the relevance of GO terms and subcellular localization. [10,12,15–19] It has been shown that GO terms can be used to advance the performance of subcellular localization prediction. These functional data regarded as domain knowledge are highly accurate and reliable, but this approach requires manual verification of each annotation and cannot be applied to the whole new protein; therefore it is usually combined with the homology-based approach. The optimized combined approach can greatly increase the predictive accuracy. 3) Sequence-based prediction relies on information about the primary amino acid sequence of proteins used for information technology operations or discovery of hidden information, commonly including amino acid composition, pseudo amino acid composition, and n-grams [19–22]. Prediction results from this approach are typically less informative than those for homology- and functional domain-based methods, but in predicting subcellular localization of unknown proteins, it is still a feasible approach. There are also systems based on n-grams for processing the amino acid sequence to analyze the subcellular location of the target protein [21]. Most of the existing methods described above only predict singleplex proteins, those that localize to a single subcellular location, and they do not consider multiplex proteins, those that localize to two or more different compartments within the cell. Identifying multiple locations of a multiplex protein has high value to understanding biological functions, and there is much room for continued development in this area [23]. Currently, the development of subcellular location prediction tools faces two major problems: the difference among the data quantity of locations is too large and the poor predictive ability for multiplex proteins [24]. Therefore, a highly accurate system was developed for predicting human protein subcellular localization that we called REALoc (Reliable and Effective methods to Assist predicting human protein subcellular Localization), which consists of two systematic frameworks. The first level uses 32 support vector machine (SVM) models in a oneto-one relationship system to completely cover all negative learning information by comprehensive strategy; the second level uses a many-to-many relationship system comprised of two learning mechanisms, a genetic algorithm optimized neural network (GANN) and majority voting (Vote), which represent the strong and weak data correlation, respectively. REALoc can predict localization to the cell membrane, cytoplasm, endoplasmic reticulum/Golgi, mitochondrion, nucleus, and extracellular. With regards to the learning sequences, this study adopted
PLOS ONE | https://doi.org/10.1371/journal.pone.0178832 June 28, 2017
2 / 14
Realoc
not only the commonly used amino acid composition (AAC) and surface accessibility (SA), but also two new learning features, weighted sign amino acid index (aa index) and sequence similarity profile. Weighted sign aa index allows more emphasis to be placed on subcellular localization-related amino acid features in the learning process, while sequence similarity profile provides identifiable learning information for similar sequences in proteins with different subcellular localization. In regards to structural learning features, we used the regular maximum relevance minimum redundancy (regular-mRMR) method to select the 35 best localization- related sets of features for GO. An independent testing dataset was used to verify the predictive power of REALoc, and the prediction results were compared to six current predictors, CELLO [25], locTree2 [26] iLochum [14], Hum-mPLoc 2.0 [27], GOASVM [6] and mGOF-loc [22]. REALoc also uses the absolute true success rate (ATSR) [9,14,24] to represent the protein prediction score, which avoids under- and over-prediction. Using the systematic ATSR evaluation method we objectively evaluated the reliability of the human subcellular localization prediction system. After five-fold cross-validation, the ATSR with the training data and independent testing data used in this study were 75.3 and 56.5%, respectively. Whether a protein was located in a single or multiple subcellular compartments, REALoc showed higher overall performance compared with other prediction systems, with a >10% improvement in the ATSR in independent testing. In terms of the analysis of the REALoc system, we also removed the GO information to evaluate the predictive power for novel proteins, and the prediction efficacies of the Vote and GANN learning mechanisms for singleplex and multiplex protein data were also analyzed.
Materials and methods Data processing Training dataset 5939p. The dataset was obtained from the UniProtKB/Swiss-Prot protein database October 2011 version [28]. Proteins with a comment on subcellular location were extracted and those with uncertain terms such as “by similarity,” potential,” and “probable” were removed. Sequences
